[MPICH] Disconnecting spawned processes after MPI_Intercomm_merge

[Rajeev Thakur]

> The problem occurs when I create an intracommunicator with 
> MPI_Intercomm_merge(). I've tried freeing and disconnecting 
> from the new 
> intracomm but the child processes seem to be waiting for the 
> parent to MPI_Finalize(). 
 
It's a bug in the current release of MPICH2 that the finalize on the client
synchronizes with the finalize on the server. It happens even without the
intercomm_merge I think.

Rajeev


> -----Original Message-----
> From: owner-mpich-discuss at mcs.anl.gov 
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of James S Perrin
> Sent: Tuesday, September 26, 2006 12:03 PM
> To: mpich
> Subject: [MPICH] Disconnecting spawned processes after 
> MPI_Intercomm_merge
> 
> Hi,
> 	I'm following up to my earlier query about using forked 
> processes in MPI . As 
> this was considered a no-no I've being trying to put together 
> the appropiate 
> framework using spawned processes. Simply I want to start a 
> set of spawned 
> processes and then end them (due to a 3rd party API forcing 
> an exit) and keep 
> the parent processes running.
> 
> 	After searching online and looking at the MPI-2 spec it 
> seems the use of 
> MPI_Comm_disconnect is not well documented. However I've got 
> a simple example to 
> work, which just launches and closes the spawned processes once.
> 
> 	The problem occurs when I create an intracommunicator with 
> MPI_Intercomm_merge(). I've tried freeing and disconnecting 
> from the new 
> intracomm but the child processes seem to be waiting for the 
> parent to 
> MPI_Finalize(). How do I release the intracomm? Is this a bug 
> or feature?
> 
> 	"Using MPI" implies that I might be better off using 
> intercommunicators for my 
> "parallel-server, parallel-client" application so I guess I 
> should explore this.
> 
> spawn_master.c
> int main (int argc, char * argv[])
> {
>    int no_exit = 1;
>    MPI_Comm inter_comm, new_world;
> 
>    MPI_Init (&argc, &argv);
>    MPI_Comm_spawn("./spawn_worker", &argv[1], 1, MPI_INFO_NULL, 0, 
> MPI_COMM_WORLD, &inter_comm, errs);
>    MPI_Intercomm_merge(inter_comm, 0, &new_world);
> 
>    // Let worker do its thing
> 
>    MPI_Comm_free(&new_world);
>    MPI_Comm_disconnect(&inter_comm);
> 
>    while(no_exit)
>    {
>       // go and do something else
>    }
> 
>    MPI_Finalize();
>    return EXIT_SUCCESS;
> }
> 
> 
> spawn_worker.c
> int main (int argc, char * argv[])
> {
>    MPI_Comm parent, new_world;
> 
>    MPI_Init (&argc, &argv);
>    MPI_Comm_get_parent(&parent);
>    MPI_Intercomm_merge(parent, 1, &new_world);
> 
>    // Do something useful
> 
>    MPI_Comm_free(&new_world);
>    MPI_Comm_disconnect(&parent);
>    MPI_Finalize();
>    return EXIT_SUCCESS;
> }
> 
> 
> Appologies for rambling but I kept coming across new bits of 
> info as I was 
> writing this.
> 
> Regards
> James
> -- 
> --------------------------------------------------------------
> ---------------
> James S. Perrin,                   | email: 
> james.perrin at manchester.ac.uk
> Manchester Visualization Centre,   | 
> www.sve.man.ac.uk/General/Staff/perrin
> Kilburn Building, The University,  | tel: +44 161 275 6945
> Manchester, England. M13 9PL.      | fax: +44 161 275 6800/6040
> --------------------------------------------------------------
> ---------------
> "The test of intellect is the refusal to belabour the 
> obvious" -Alfred Bester
> --------------------------------------------------------------
> ---------------
> 
>